Update README.md
Browse files
README.md
CHANGED
|
@@ -10,13 +10,40 @@ license: other
|
|
| 10 |
|
| 11 |
# VLAC2 Benchmark Release
|
| 12 |
|
| 13 |
-
This release
|
| 14 |
|
| 15 |
-
|
| 16 |
-
- `scripts/`: frame extraction utilities
|
| 17 |
-
- `vlac2_release_data/`: raw-data archives plus their manifest
|
| 18 |
-
- `docs/`: minimal usage notes
|
| 19 |
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
|
| 11 |
# VLAC2 Benchmark Release
|
| 12 |
|
| 13 |
+
VLAC2 is a video-progress benchmark for long-horizon robot manipulation. This release provides benchmark JSON files with dense ground-truth progress targets, the associated raw-data archives, and a minimal frame-extraction workflow for reproducible evaluation.
|
| 14 |
|
| 15 |
+
## Contents
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
+
- `benchmark_json/train/video_progress_benchmark_file.json`
|
| 18 |
+
- `benchmark_json/test_expert_seen/video_progress_benchmark_file.json`
|
| 19 |
+
- `benchmark_json/test_expert_unseen/video_progress_benchmark_file.json`
|
| 20 |
+
- `benchmark_json/test_nonexpert_seen/video_progress_benchmark_file.json`
|
| 21 |
+
- `benchmark_json/test_nonexpert_unseen/video_progress_benchmark_file.json`
|
| 22 |
+
- `scripts/extract_vlac2_release_raw_archives.sh`
|
| 23 |
+
- `scripts/extract_vlac2_release_frames.py`
|
| 24 |
+
- `scripts/vlac2_release_common.py`
|
| 25 |
+
- `vlac2_release_data/vlac2_release_data_part*.tar.zst`
|
| 26 |
+
|
| 27 |
+
## Usage
|
| 28 |
+
|
| 29 |
+
1. Unpack all raw-data archives into `vlac2_release_data/`.
|
| 30 |
+
|
| 31 |
+
```bash
|
| 32 |
+
bash scripts/extract_vlac2_release_raw_archives.sh /path/to/vlac2_release/vlac2_release_data
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
2. Extract benchmark frames from the raw videos.
|
| 36 |
+
|
| 37 |
+
```bash
|
| 38 |
+
python scripts/extract_vlac2_release_frames.py --release-root /path/to/vlac2_release
|
| 39 |
+
```
|
| 40 |
+
|
| 41 |
+
If `--output-root` is not specified, extracted frames are written to `/path/to/vlac2_release/_extracted_frames/`, which is created automatically by the script.
|
| 42 |
+
|
| 43 |
+
3. Use the benchmark JSON files for training or evaluation. Frame paths in the JSONs use the placeholder prefix `__VLAC2_RELEASE_ROOT__`, which should be replaced with the absolute path of the local release root.
|
| 44 |
+
|
| 45 |
+
## Notes
|
| 46 |
+
|
| 47 |
+
- The benchmark JSON files are ready to use; users do not need to rebuild them.
|
| 48 |
+
- A video decoder backend is required for frame extraction: `opencv-python` or `imageio`.
|
| 49 |
+
- The raw-video distribution and the extracted-frame cache are intentionally separated so that users can regenerate frames locally under their own storage layout.
|